On 02/03/2017 01:12 PM, Segher Boessenkool wrote:
Scheduling should never move very expensive instructions to places they
are executed more frequently. This patch fixes that, reducing the
execution time of c-ray by over 40% (I tested on a BE Power7 system).
This introduces a new target hook sched.can_speculate_insn which returns
whether the scheduler is allowed to speculate a given instruction. The
rs6000 implementation disallows all divide and square root instructions.
Bootstrapped and tested on powerpc64-linux {-m32,-m64}. Is this okay
for trunk?
Segher
2017-02-03 Segher Boessenkool <seg...@kernel.crashing.org>
PR rtl-optimization/68664
* target.def (can_speculate_insn): New hook.
* doc/tm.texi.in (TARGET_SCHED_CAN_SPECULATE_INSN): New hook.
* doc/tm.texi: Regenerate.
* sched-rgn.c (can_schedule_ready_p): Use the new hook.
* config/rs6000/rs6000.c (TARGET_SCHED_CAN_SPECULATE_INSN): New macro.
(rs6000_sched_can_speculate_insn): New function.
---
gcc/config/rs6000/rs6000.c | 20 ++++++++++++++++++++
gcc/doc/tm.texi | 6 ++++++
gcc/doc/tm.texi.in | 2 ++
gcc/sched-rgn.c | 19 +++++++++++++------
gcc/target.def | 7 +++++++
5 files changed, 48 insertions(+), 6 deletions(-)
diff --git a/gcc/doc/tm.texi b/gcc/doc/tm.texi
index 909589c..1963402 100644
--- a/gcc/doc/tm.texi
+++ b/gcc/doc/tm.texi
@@ -7000,6 +7000,12 @@ The structure *@var{spec_info} should be filled in by
the target.
The structure describes speculation types that can be used in the scheduler.
@end deftypefn
+@deftypefn {Target Hook} bool TARGET_SCHED_CAN_SPECULATE_INSN (rtx_insn
*@var{insn})
+Some instructions should never be speculated by the schedulers, usually
+ because the instruction is too expensive to get this wrong. This hook
+ should return @code{false} if @var{insn} should not be speculated.
+@end deftypefn
Consider adding something like this:
Define this hook to return false for instructions which are not fully
modeled by the pipeline description to avoid DFA size explosion.
Otherwise the scheduler may erroneously speculate those instructions
into a pipeline bubble that is too small which may severely impact
performance.
+
@deftypefn {Target Hook} int TARGET_SCHED_SMS_RES_MII (struct ddg *@var{g})
This hook is called by the swing modulo scheduler to calculate a
resource-based lower bound which is based on the resources available in
diff --git a/gcc/doc/tm.texi.in b/gcc/doc/tm.texi.in
index ea74d37..756c118 100644
--- a/gcc/doc/tm.texi.in
+++ b/gcc/doc/tm.texi.in
@@ -4882,6 +4882,8 @@ them: try the first ones in this list first.
descr
@hook TARGET_SCHED_SET_SCHED_FLAGS
+@hook TARGET_SCHED_CAN_SPECULATE_INSN
+
@hook TARGET_SCHED_SMS_RES_MII
@hook TARGET_SCHED_DISPATCH
diff --git a/gcc/sched-rgn.c b/gcc/sched-rgn.c
index 2af3a03..a09fc5d 100644
--- a/gcc/sched-rgn.c
+++ b/gcc/sched-rgn.c
@@ -2147,12 +2147,19 @@ static int
can_schedule_ready_p (rtx_insn *insn)
{
/* An interblock motion? */
- if (INSN_BB (insn) != target_bb
- && IS_SPECULATIVE_INSN (insn)
- && !check_live (insn, INSN_BB (insn)))
- return 0;
- else
- return 1;
+ if (INSN_BB (insn) != target_bb && IS_SPECULATIVE_INSN (insn))
+ {
+ /* Cannot schedule this insn unless all operands are live. */
+ if (!check_live (insn, INSN_BB (insn)))
+ return 0;
+
+ /* Should not move expensive instructions speculatively. */
+ if (GET_CODE (PATTERN (insn)) != CLOBBER
+ && !targetm.sched.can_speculate_insn (insn))
+ return 0;
+ }
You probably want to filter both CLOBBER and USE insns.
OK with those two nits addressed.
jeff
